introduction to time-series analysis with pi system data...yeast can easily form a membrane that...

#PIWorld ©2018 OSIsoft, LLC

Introduction to Time-Series Analysis with PI System Data

Prof. Brian D. Davison

Prof. Martin Takáč



Presentation Outline

• Preliminaries: Sensors, Data set, and Sampling

• Task #1: Prediction of ADF

• Task #2: Modeling Cooling

• Task #3: PCA for Visualization

• Lessons Learned


3

Preliminaries Sensors, Data set, and Sampling


Sensor data

4

Common problems:

• missing values (wrong reading from sensor/communication issue)

• a sensor fails and gives wrong readings

• frequent readings of sensor values (e.g., each second/minute)


Brewery Dataset

5

source: wikipedia

Fermenter

Temperature

sensors top

middle

bottom

Pressure

sensor

… and many more sensors…

and even more fermenters...


Brewery Dataset

• From a company that produces more than two dozen craft beers

• Different beers can use different yeasts during fermentation and different fermenting temperatures

6

Stages of beer fermentation

From OSIsoft Users Conference 2017

OSIsoft: Introduction to Process

Optimization and Big-Data Analytics

with the PI System

https://www.youtube.com/watch?v=9Ad

vKkyTeBE

https://www.youtube.com/watch?v=9AdvKkyTeBE

https://www.youtube.com/watch?v=9AdvKkyTeBE


Formally, a time series

7

• is an ordered sequence of values of a variable at equally spaced time intervals

• can be built on top of data obtained from sensors

– choose the size of time interval (e.g., 1 hour)

– represent all measurements in each hour by one value (e.g., average value / first value / last value …)

– we obtain a collection of values (e.g., one number per hour)

– the PI system can extract such values from collected data

• has many applications

– in forecasting, monitoring, or feedback and feed-forward control


Multiple sources and “sampling” frequencies

8

Real life problem = multiple sensors

Manual = infrequent


An illustrative Example

9

bottom

middle

top temperature sensor

Before doing analytics, we need to align the data streams (in this example, we could choose a multiple of one hour)!


10

Task #1 Prediction of ADF


The Problem Setting

11

• Automated measuring of gravity of wort is expensive

• The brewery measures it manually (and infrequently)

• It is used to compute Apparent Degree of Fermentation (ADF)

which is an important indicator of the status of fermentation

(Note: ADF indicates how much sugar has been converted to alcohol)

Our task: Predict ADF as well as possible eliminate a lot of manual work

improve quality of beer

decrease cost of fermentation


Data and Challenges

12

• We can use the data from all sensors to predict ADF BUT

• We have only a few ADF values per fermentation cycle (and moreover it is not computed in “regular” intervals)


The First Approach

13

• Let’s fix one beer brand, and plot our ADF points: x-axis - Time from start of fermentation y-axis - ADF

MyData <- read.csv(file="regression.txt")

MyData$time <- MyData$time/60/60

plot(MyData$time, MyData$adf, xlab =

"Time Since Fermentation [hours]",

ylab = "ADF", main = "Data without outliers")


Simple Linear Regression Model

14

response (dependent)

variable

explanatory (independent)

variable

intercept slope

errors of the

response

variables should

be uncorrelated

with each other


Linear Fit

15

Estimate

(Intercept) -0.0598497

Time 0.0159091

each hour, ADF

increases by

~0.016

plot(FilteredTimes, FilteredADF, xlab = "Time Since

Fermentation [hours]", ylab='ADF', main="ADF ~

Time")

linMod <- lm(FilteredADF ~ FilteredTimes)

abline(linMod$coefficients)

summary(linMod)


Checking Assumptions with a Residual Plot

16

errors should be

independent of

predicted values

Our model is NOT ok!

predicted <- fitted(linMod)

errors <- resid(linMod)

plot(predicted, errors, main = "Residuals vs. predicted

ADF")


What was wrong? How to improve it?

17

• We tried our luck with a simple linear model

• The model works OK for the data in the middle of our range, but is failing for small and large values of ADF

• As data scientists, we need to understand our problem domain.

• We need to learn about beer fermentation!


Understanding the Fermentation Process (a bit better)

18

• What influences the speed of fermentation?

– temperature (we are cooling the wort, so this is relatively constant)

– amount of yeast

– amount of sugar and ethanol level

Gilliland, R.B., 1962. Yeast

reproduction during fermentation.

Journal of the Institute of Brewing,

68(3), pp.271-275.


Speed of Fermentation

19

time http://wineserver.ucdavis.edu/industry/enology/fermentation_management/wine/problem_fermentations.html

sugar

content

One degree Brix

is 1 gram of

sucrose in 100

grams of solution

cells are adapting to the wort conditions and

engaging in cell division

fastest rate of fermentation, ethanol

level is relatively low

The rate slows because ethanol in

the medium forces an adaptation

of the plasma membrane. The

yeast can easily form a

membrane that depends upon

ethanol replacing water for its

structure and functionality…

http://wineserver.ucdavis.edu/industry/enology/fermentation_management/wine/problem_fermentations.html


New (improved) model

20

• We now understand that there are different rates of fermentation.

• How about fitting 3 linear models? We would also need to figure out the break-points.

parameters to estimate


intercept above zero

initial slope

Fitted Model

21

increase of slope: new

slope is 0.018011

decrease of slope: new

slope is 0.003764

break-points

xx <- FilteredTimes

yy <- FilteredADF

dati <- data.frame(x = xx, y = yy )

out.lm <- lm(y ~ x, data = dati)

o <- segmented(out.lm, seg.Z = ~x,

psi = list(x = c(10,40)),

control = seg.control(display = FALSE))

dat2 = data.frame(x = xx, y = broken.line(o)$fit)

o

Meaningful coefficients of the linear terms: (Intercept) x U1.x U2.x 0.005924 0.009560 0.008451 -0.014247 Estimated Break-Point(s): psi1.x psi2.x 14.57 45.61


Fitted Model

22

ggplot(dati, aes(x = x, y = y)) + xlab("Time Since

Fermentation [hours]") +

ylab('ADF') +

geom_point() +

geom_line(data = dat2, color = 'blue')


Checking Assumptions with a Residual Plot

23

errors <- resid(o)

plot( predicted, errors, main = "Residuals vs. predicted

ADF" )


24

Task #2 Cooling prediction


The Problem Setting

25

• After the fermentation (and possibly Free Rise or Diacetyl Rest) is finished, the beer has to be cooled almost to a freezing point

• The data contains temperature measurements at 3 different locations

Top

Middle

Bottom


Typical Cooling Profile

26

In the ideal case, the

beer is cooling nicely

and all sensors read

similar values


Atypical Cooling Profile

27

Top

Middle

Bottom


All together

28


Issues with data

29

• For each time point, we know in

which state of fermentation the

beer is (colors on graph)

• We also know about other sensors,

such as the cooling valves

• During fermentation, we try to

keep temperature constant

• The cooling is achieved by three

cooling valves (usually open 100%)

• However, there are many inconsistencies

Fermentation

Free rise

Diacetyl

Rest Cooling


Issues with data

30

Cooling

labeled as

Diacetyl Rest


Issues with data

31


• We will ignore the “label” of the fermentation “stage”

• We define a cooling period simply as a uninterrupted interval with 100% cooling capacity

• Moreover, we will be interested in the data until the temperature reaches ~32F

Preprocessing

32

Free rise Cooling


Filtered cooling profiles

33

● different initial

temperature for

various coolings


Volumes

34

● various volumes of

beer in fermenter


Thermodynamic heat transfer

35

• how the temperature will change in a small time interval ?

• real answer is hard and involves solving PDFs, …..

• we can try to approximate it

New temperature

Old temperature

Temperature

of “cooling”

substance

Cooling Rate ( <0 ) Volume


Some simplifications

36


Time difference = 30 min

37

Parameters:

Estimate Std. Error t value Pr(>|t|)

alpha -17.1918 0.3134 -54.86 <2e-16 ***

beta 30.4543 0.2427 125.49 <2e-16 ***

The cooling

substance is

~30.4F

d <- read.csv(file="UC2_TS_30.dat")

yOld <- d$o1

y <- d$n1

vol <- d$v

m <- nls(y ~ (1+a/vol)*yOld - a*b/vol)

summary(m)


Residuals

38

we have more data when it is cooler

predicted <- fitted(m)

errors <- resid(m)

plot(predicted, errors) Residuals vs. Predicted Temperature


Real cooling profile vs. Predicted cooling profile

39

real, measured data

computed from initial

temperature and volume

nd <- read.csv("UC2_Prediction.dat")

x<-seq(0,58*4,1)*0.25; y<-x*0

y[1] <- nd$m1[1]; vol <- nd$v[1]

alpha <- summary(m)$coefficients[1]

beta <- summary(m)$coefficients[2]

for (i in 2:length(x)) {

y[i] <- (1+alpha/vol)*y[i-1] - alpha*beta/vol

}

plot(nd$t/60, nd$m1, xlab = "Time Since Cooling [hours]",

ylab='Temperature [F]', type='l', col="red")

lines(x, y, type="s", col="blue")


40

Task #3 Principal component analysis for

visualization


Principal component analysis (PCA)

41

• PCA performs dimensionality reduction and de-correlation of dimensions

• The “real” data can lie within a smaller dimensional subspace

• PCA is useful for visualization and pre-processing

• Detect patterns and outliers

• Produce fewer and uncorrelated dimensions

• Good for data with relatively few labels


Data preprocessing

42

• We have a lot of data from sensors; it is a bit tricky to use PCA directly (too many points, missing data)

• More importantly, we want to see when cycles might be same or differ

• We represented each fermentation cycle by 31 features, like

– highest ADF

– mean/min/max temperature during fermentation

– volume

– mean/min/max temperature during cooling

– duration of fermentation

– duration of cooling

– …...


Visualization in 3D

43

how much variance is captured by the

ith principal component • We scale the data and compute the principal components

• We project the data onto two or three principal components that represent the most variance and plot them

# apply PCA - scale. = TRUE is highly

# advisable, but default is FALSE.

beers.pca <- prcomp(beers.data,

center = TRUE,

scale. = TRUE)

plot(beers.pca, type = "l")


Visualization in 3D

44

• "Realtime Hops"

• "Red Wonder"

• "Alistair"

• "Kennedy"

• "Kerberos"

# Predict PCs

beers.projected <- predict(beers.pca,

newdata=beers.data)

scatterplot3d(beers.projected[,1],

beers.projected[,2],

beers.projected[,3],

color = beers.labels,

main = "PCA")


Similar members

45


Outliers

46


47

Lessons Learned


RESULTS CHALLENGE SOLUTION

COMPANY and GOAL

48

Lesson Learned

Predicting Apparent Degree of Fermentation (ADF)

Linear model of ADF as a function of time is insufficient

Use domain knowledge to develop a better model

Piecewise linear model is much improved over basic linear model

• Existence of three phases of fermentation suggests three segments in the linear model

• Errors are not correlated with predicted values

• Normality test values are better (not presented)

• Could consider other attributes

• ADF starts negative

• Errors are correlated with predicted values (BAD!)

• Normality test values are not near model (not presented)



COMPANY and GOAL

49

Lesson Learned

Predict cooling of fermented beer

Want to predict cooling trend, completion times

Basic thermodynamics function for cooling

Basic cooling model works well

• Use data to find values of constants

• Errors are not correlated with predicted values

• Could build complex model

• Many attributes available, including starting temperature, volume, and time



COMPANY and GOAL

50

Lesson Learned

Visualizing high-dimensional data

Want to see similarities and differences between runs

Use PCA on run-focused data Can see outliers and patterns

• Represent situation of interest

• Identify reduced dimensionality

• Some beers are consistent

• Others include outliers

• Data in form of time series with many attributes

• Humans need at most 3-D data


References for further study

• Gebhard Kirchgässner and Jürgen Wolters, Introduction to Modern Time Series Analysis, Springer Berlin Heidelberg New York, 2007.

• Peter J. Brockwell and Richard A. Davis, Introduction to Time Series and Forecasting, 2nd ed., Springer Texts in Statistics, Springer Science+Business Media, LLC, New York, NY, USA, 2002.

• James D. Hamilton, Time Series Analysis, Princeton University Press, Princeton, NJ, USA, 1994.

• NIST/SEMATECH e-Handbook of Statistical Methods, http://www.itl.nist.gov/div898/handbook/, 2018.

• http://www.r-tutor.com

• http://wineserver.ucdavis.edu/industry/enology/fermentation_management/wine/problem_fermentations.html

51

http://www.itl.nist.gov/div898/handbook/

http://www.r-tutor.com/






Contact Information

Brian D. Davison [email protected]

Associate Prof., Lehigh University

Slides and R source code available from:

http://www.cse.lehigh.edu/~brian/PIW18/ Lehigh University's Linderman Library

http://www.cse.lehigh.edu/~brian/PIW18/


Questions?

Please wait for

the microphone

State your

name & company

Please rate this session

in the mobile app!

Search

“OSIsoft” in

your app store

53


54

introduction to time-series analysis with pi system data...yeast can easily form a membrane that...

Documents